plot

Plot

Drop-in matplotlib templates for publication-quality figures. Each template is one .py file producing one figure, with no save logic — copy, swap data, add your savefig.

Principles

The complete style contract. Every figure — template-derived or written from scratch — follows all of these; style.py encodes most of them as rc defaults and helpers.

1. Frame: announcement-clean. L-shaped spines only (top/right hidden), near-black ink #1a1a1a for spines/ticks/labels, no grid, outward ticks. All rc defaults — never restate them in a script; clean_axes(ax) re-asserts the frame only on twin/secondary axes the rc cannot reach.
2. Titles: left-aligned, bold, no letter prefixes. Plain ax.set_title(text) is enough — the rc makes it left-aligned bold ink, and titles are the ONLY bold text in a figure. Never prefix titles with (a)/(b)/(c) — reference panels in captions as Left/Middle/Right or Top/Bottom. Multi-panel suptitles: fig.suptitle(..., x=0.01, ha='left', fontweight='bold', color=INK).
3. Legend: a header row above the axes, never inside them. header_legend(ax, entries) per axes, or fig_header_legend(fig, entries) for one figure-level row over a multi-panel grid (needs constrained_layout). Entries: (label, color) for white-edged dots, (label, color, '-') solid-line proxy, (label, color, '--') reference-dash proxy, or any marker char. Never hand-roll proxy handles; keep labels short enough that the row fits its panel's width.
4. Spacing: measured, not guessed. End every figure with finalize_headers(fig) — after all set_title/header_legend calls, before savefig. It measures the real legend heights and makes title, legend, and plot equidistant (one 6pt gap on each side of the legend row), level across panels, at any font size. level_all=False when legend-less panels sit in their own row under a figure-level header. After ANY header or font change, re-render and view the output — never ship a spacing change unchecked.
5. Palette: Google brand only. G_BLUE/G_RED/G_YELLOW/G_GREEN/G_PURPLE/G_GREY softened through the tier system (brand → medium → paper (default) → soft → mute). Paper series colors default to the medium tier.
6. Hue count follows series count. At most 3 colored series → ONE brand hue, lightness steps via hue_ramp(base, n) (index 0 lightest) or twotone(base). More than 3 series → distinct Google hues at the medium tier. Neutrals never count as a hue: HUMAN_DARK/HUMAN_SOFT for human/reference cohorts, greys for annotations. Legend-less encodings (bar charts read off the axis) may use longer ramps.
7. References: one grey, one dash. Every reference/baseline line is REF_GREY + REF_DASH. A second dashed series in the same panel is distinguished by color and label, reusing REF_DASH rather than inventing a new pattern.
8. Markers and bands. Emphasized markers get markeredgecolor='white', width 0.6–0.8. Confidence bands are the same hue as their line, alpha 0.12–0.18, linewidth=0.
9. Fonts: Palatino with real bold, one size band. Palatino body + STIX math (matches LaTeX mathpazo); apply_style() registers TeX Gyre Pagella from TeX Live because macOS Palatino.ttc exposes no bold face to matplotlib. Sizes: title 12.5 bold / axis label 14 / tick 13 (rc), header legend rows 9.5, annotations ≥ 8.5. Dense multi-panel grids may step down (title ~10, tick ~8), and every figure in one document stays inside this one band.
10. Output: PDF is the artifact. Save both out.pdf and a dpi=200 PNG preview; papers include only the PDF. matplotlib.use('Agg') before pyplot in scripts; anchor outputs on HERE = Path(__file__).resolve().parent.
11. Restyle never touches data. A styling pass changes colors, legends, fonts, and spacing — not data loading, fits, tick semantics, or panel content. After restyling a figure with computed values, verify the numbers are identical to the pre-restyle run.